DAM-Tracker: A Real-Time Siamese Network Tracking Algorithm via Deep and Dynamic Attention Collaboration

Visual Object Tracking (VOT) remains a fundamental challenge in computer vision. In recent years, deep learning algorithms such as Convolutional Neural Networks (CNNs) and Transformers have emerged as dominant solutions, demonstrating significant performance improvements across benchmark datasets. Notably, Transformer architectures based on attention mechanisms have been widely adopted, particularly when trained on large-scale datasets. However, existing trackers often exhibit critical limitations: (1) insufficient integration of local and global features, (2) neglect of spatiotemporal information extraction and utilization, and (3) inherent feature redundancy in vanilla Transformer structures, which introduces extraneous signal interference and increases susceptibility to irrelevant regions, thereby compromising tracking accuracy and efficiency. To address these issues, we propose DAM-Tracker, a real-time Siamese network tracking algorithm based on deep and dynamic attention collaboration. Extensive experiments on the OTB100 and UAV123 benchmarks validate its superior performance: 1) DAM-Tracker achieves competitive tracking results in challenging sequences and complex scenarios compared to state-of-the-art trackers; 2) The simplified architecture maintains high accuracy while achieving an average tracking speed of 35 frames per second (FPS).